3 ms·

This sounds intriguing and slightly hype-y. But I would love to see our math guys here break down and analyze what's hype and what's not. Is Claude's descriptio

This sounds intriguing and slightly hype-y. But I would love to see our math guys here break down and analyze what's hype and what's not. Is Claude's description detailed and concrete enough to be useful? It seems that's where LLM's often struggle.

onlyrealcuzzo 3y ago
It's probably been prompted like this for similar break-throughs thousands of times. Is it really that impressive that it's done this once? I imagine with enough tries, GPT-4 Turbo and Gemini 1.5 Pro have done similarly impressive things.
- bigyikes 3y ago
  Let it be known: this is what moving the goalposts sounds like! The AI is proposing novel quantum algorithms using terminology most of us here don’t understand, and it’s apparently not “that impressive”! We truly live in interesting times
  - mewpmewp2 3y ago
    True, with AI it feels like some people are overprotective of their egos, having to downplay the amazing achievements we have seen since end of 2022. Maybe they are subconsiously so afraid of actual AGI coming, that they do have to ignore the achievements as of now.
  - skybrian 3y ago
    If you were previously talking with someone who thought that AI's can't be creative and now they say it doesn't matter, then some goalposts moved. More likely, though, you're just reading commentary by someone else who has different goalposts in mind. Unlike football, there is no standard for goalposts in AI.
  - nicklecompte 3y ago
    The reason it's not that impressive is a lot of us are not as easily fooled as we might have been back in 2022! As someone who also went to UWaterloo for quantum computing (MMath, not PhD, and I don't work in the field), let's look at the tweet in detail: - in the first image, Claude's output is almost all useless LLM mush! Points #1, #3, #4, #5, #6, and #7 are uselessly vague and at best tangentially related to the problem. Only #2 has anything of interest, and even then it's just "come up with a quantum version of a classical algorithm," which is hardly insightful. Claude's output here just straight-up sucks! - in the second image, "Initialization" is useless, "Quantum State Measurement" is useless, "Momentum Refreshment" is useless (and might be wrong in this context), "Iterations" is useless and misleading, and "Sample Collection" is useless. - There is some real meat in the "Quantum Leapfrog Integration" section, but the entire last bullet point (starting with "The quantum leapfrog integrator consists of") is a completely useless statement which applies to every quantum algorithm. There is almost nothing actionable for a quantum computing researcher here, except maybe as a brainstorming aid. And of course the giant pile of LLM mush is evidence that Claude doesn’t know what a “quantum computer” is or what an “algorithm” is, it’s just a fancy chatbot. I am not at all impressed by any of this.
  - jprete 3y ago
    If you ask for a thousand novel theories, and all the answers look good but only five can possibly work, you haven't done much better than talking to physics cranks.
amluto 3y ago
It reads like something between gobbledygook and an executive summary of a paper with maybe some potentially relevant fancy terms thrown in for good measure. I have no idea how good the actual paper is, but the Claude output would not make an impressive paper. (It’s still pretty neat that an LLM invented it with a minimal prompt!) (I’ve done quantum computing research and, while I did share an office and a research group with people doing Monte Carlo, I didn’t do it myself.)
- winwang 3y ago
  Also did a bit of QC research. Since I'm tipsy, I skimmed the Twitter post. Here's a pdf on arxiv about HMC (2017) which I found in 20 seconds: https://arxiv.org/abs/1701.02434 https://arxiv.org/abs/1701.02434 Seems like Claude3 is decently logical in paraphrasing key parts of text in a good order, but with repetition scattered everywhere. Also Monte Carlo is something which sounds cool but is literally just conceptual dart-throwing. I've always hated the name.
  - amluto 3y ago
    However you feel about the term Monte Carlo, MCMC (Markov chain Monte Carlo) is a nifty algorithm. You’re also welcome to say “Metropolis-Hastings algorithm” (assuming that’s the specific algorithm you’re using). I was personally entertained when I learned that the algorithm formerly known as the Metropolis algorithm had nothing to do with cities. And MCMC really is like throwing darts. Except that, when done right, it throws darts in a distribution of your choice, and, critically, you don’t need to be able to normalize the distribution or integrate it to find the CDF. This turns out to be extremely useful. So you can think of it as a sort of bizarre way to numerically integrate gnarly functions. If you already knew the inverse CDF of a distribution, you could directly sample it. But if you don’t, either because it’s hard to integrate or because you’re getting the probabilities or probability densities, unnormalized, from a black box, you can still run MCMC. I bet you could have some fun taking a small ML model, using it to compute probabilities of strings of tokens, and using MCMC to sample the distribution of likely strings as seen by the model. The challenging part will be coming up with a remotely credible random walk that efficiently explores the space. NB: everyone loves saying “log-likelihood”. This is essentially the same thing. The log is nice because otherwise the numbers end up awkwardly small.

onlyrealcuzzo 3y ago

It's probably been prompted like this for similar break-throughs thousands of times. Is it really that impressive that it's done this once? I imagine with enough tries, GPT-4 Turbo and Gemini 1.5 Pro have done similarly impressive things.

bigyikes 3y ago
Let it be known: this is what moving the goalposts sounds like! The AI is proposing novel quantum algorithms using terminology most of us here don’t understand, and it’s apparently not “that impressive”! We truly live in interesting times
- mewpmewp2 3y ago
  True, with AI it feels like some people are overprotective of their egos, having to downplay the amazing achievements we have seen since end of 2022. Maybe they are subconsiously so afraid of actual AGI coming, that they do have to ignore the achievements as of now.
- skybrian 3y ago
  If you were previously talking with someone who thought that AI's can't be creative and now they say it doesn't matter, then some goalposts moved. More likely, though, you're just reading commentary by someone else who has different goalposts in mind. Unlike football, there is no standard for goalposts in AI.
- nicklecompte 3y ago
  The reason it's not that impressive is a lot of us are not as easily fooled as we might have been back in 2022! As someone who also went to UWaterloo for quantum computing (MMath, not PhD, and I don't work in the field), let's look at the tweet in detail: - in the first image, Claude's output is almost all useless LLM mush! Points #1, #3, #4, #5, #6, and #7 are uselessly vague and at best tangentially related to the problem. Only #2 has anything of interest, and even then it's just "come up with a quantum version of a classical algorithm," which is hardly insightful. Claude's output here just straight-up sucks! - in the second image, "Initialization" is useless, "Quantum State Measurement" is useless, "Momentum Refreshment" is useless (and might be wrong in this context), "Iterations" is useless and misleading, and "Sample Collection" is useless. - There is some real meat in the "Quantum Leapfrog Integration" section, but the entire last bullet point (starting with "The quantum leapfrog integrator consists of") is a completely useless statement which applies to every quantum algorithm. There is almost nothing actionable for a quantum computing researcher here, except maybe as a brainstorming aid. And of course the giant pile of LLM mush is evidence that Claude doesn’t know what a “quantum computer” is or what an “algorithm” is, it’s just a fancy chatbot. I am not at all impressed by any of this.
- jprete 3y ago
  If you ask for a thousand novel theories, and all the answers look good but only five can possibly work, you haven't done much better than talking to physics cranks.

amluto 3y ago

It reads like something between gobbledygook and an executive summary of a paper with maybe some potentially relevant fancy terms thrown in for good measure. I have no idea how good the actual paper is, but the Claude output would not make an impressive paper. (It’s still pretty neat that an LLM invented it with a minimal prompt!) (I’ve done quantum computing research and, while I did share an office and a research group with people doing Monte Carlo, I didn’t do it myself.)

winwang 3y ago
Also did a bit of QC research. Since I'm tipsy, I skimmed the Twitter post. Here's a pdf on arxiv about HMC (2017) which I found in 20 seconds: https://arxiv.org/abs/1701.02434 https://arxiv.org/abs/1701.02434 Seems like Claude3 is decently logical in paraphrasing key parts of text in a good order, but with repetition scattered everywhere. Also Monte Carlo is something which sounds cool but is literally just conceptual dart-throwing. I've always hated the name.
- amluto 3y ago
  However you feel about the term Monte Carlo, MCMC (Markov chain Monte Carlo) is a nifty algorithm. You’re also welcome to say “Metropolis-Hastings algorithm” (assuming that’s the specific algorithm you’re using). I was personally entertained when I learned that the algorithm formerly known as the Metropolis algorithm had nothing to do with cities. And MCMC really is like throwing darts. Except that, when done right, it throws darts in a distribution of your choice, and, critically, you don’t need to be able to normalize the distribution or integrate it to find the CDF. This turns out to be extremely useful. So you can think of it as a sort of bizarre way to numerically integrate gnarly functions. If you already knew the inverse CDF of a distribution, you could directly sample it. But if you don’t, either because it’s hard to integrate or because you’re getting the probabilities or probability densities, unnormalized, from a black box, you can still run MCMC. I bet you could have some fun taking a small ML model, using it to compute probabilities of strings of tokens, and using MCMC to sample the distribution of likely strings as seen by the model. The challenging part will be coming up with a remotely credible random walk that efficiently explores the space. NB: everyone loves saying “log-likelihood”. This is essentially the same thing. The log is nice because otherwise the numbers end up awkwardly small.